Spatial-Temporal Context-Aware Location Prediction Based on Bidirectional Self-Attention Network

The next-location prediction tasks get much attention because it is employed in many applications. The accuracy of location prediction has become the basis of these applications. The existing approaches related rely on transition matrices according to specific probabilistic rules or recurrent neural networks that cannot be applied to complex scenarios. Other works focus on extracting extra information in trajectory. In this paper, we propose a context-aware model with a bidirectional self-attention network for location prediction, which can capture implicit spatial-temporal patterns from the time stamps and geographical distances of locations. Besides, a training mechanism, Mask Locations, is employed to improve the prediction accuracy. We conduct experiments on two large-scale datasets: a check-in dataset and a Call Detail Record (CDR) dataset. The results show that our model significantly outperforms the competitive baseline methods.